Platform
The architecture behind autonomous trading
ArcAlpha is built as an actor system: independent agents reasoning over shared intelligence, executing through one risk-checked path, and checkpointing their state so the whole fleet survives any restart.
Architecture
One pipeline, end to end
Market data flows through intelligence and decisioning, clears every risk gate, and only then reaches a broker — with state checkpointed back on every cycle.
9
Agent states
60s
Decision cycle
N×
Horizontal scale
100%
State recovered
Autonomous Agents
One worker per asset, each running itself
Every traded asset gets its own TradeWorker — an independent actor with its own state, strategy, and decision cycle. Agents don't share a brain; they share infrastructure.
Independent decision cycle
Each worker analyzes, decides, and acts on its own ~60-second loop without blocking any other.
Persistent worker state
Position, strategy, last analysis, and the next planned action are written to MongoDB every cycle.
Competition & evolution
Workers compete in performance tournaments; capital flows to winners and weak agents are paused.
Self-healing lifecycle
A lifecycle manager and health monitor detect stuck agents and recover them automatically.
Agent decision loop
Execution
From decision to fill, safely
Approved decisions become orders only after they clear sizing, leverage, and day-trade rules — then route to the right venue.
Smart order routing
The OrderManager routes each order to the correct brokerage for its market and asset class.
Pre-trade safety
Position sizing, leverage caps, wash-trade handling, and PDT rules are enforced before any order leaves.
Mode equivalence
Simulation and live trading share 99% of the code path — only the final execution differs.
Runtime
Built to scale out, not just up
ArcAlpha runs across multiple web instances behind a load balancer, with fleet schedulers leader-locked so work never doubles.
Horizontal scale-out
Multiple instances serve traffic behind a sticky load balancer with a shared Redis Socket.IO adapter.
Leader-locked schedulers
Fleet-wide jobs elect a single leader so optimization and competition run exactly once.
Checkpoint & recover
On startup the PortfolioManager rebuilds every worker from its last checkpoint — nothing is lost.